AD-DAE: Alzheimer's Disease Progression Modeling with Unpaired Longitudinal MRI using Diffusion Auto-Encoders.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Generative models capture complex image data for disease progression modeling.
• Existing methods struggle with controllable latent spaces for longitudinal image generation without paired data.
• Limited controllability hinders accurate disease progression prediction.

Purpose of the Study:

• To introduce a conditionable Diffusion Auto-encoder for controlled generation of longitudinal disease progression images.
• To enable generation of follow-up images from prior time-points without subject-specific longitudinal data.
• To create a compact latent space for high-level semantic capture and controlled generation.

Main Methods:

• Developed a conditionable Diffusion Auto-encoder framework.
• Leveraged a compact latent space to isolate progression and subject identity.
• Applied controlled shifts to representations for generating follow-up images, guided by disease attributes and region constraints.
• Validated using image quality metrics, volumetric analysis, and downstream tasks on Alzheimer's disease datasets.

Main Results:

• The framework successfully generates realistic follow-up images reflecting disease progression.
• Controlled latent space manipulation allows for targeted generation without paired longitudinal data.
• Alzheimer's disease-specific region constraints improve generation accuracy.
• Downstream task performance demonstrates the model's effectiveness in disease progression modeling.

Conclusions:

• The proposed Diffusion Auto-encoder framework effectively models Alzheimer's disease progression.
• The approach offers enhanced controllability for longitudinal image generation.
• This method advances generative modeling for medical imaging and disease prediction.